Judah De Paula

1
Opponent Cell Model of the Primary Visual Cortex

• Judah De Paula
• November 12, 2003
• University of Texas, Austin
• Neural Networks Group

2
Overview

• Introduction
• Vision
• Color Vision
• LISSOM
• Creating LISSOM Orientation Maps
• Initial Runs
• Natural / Artificial Stimuli questions
• Future directions?

3
Vision Physiology
From (1)
4
Vision Physiology

• Rods not used in daylight.
• Fovea consists of cones.
• Cone wavelength sensitivity functions overlap
  each other.
• Only Long (Red) and Medium (Green) wavelength
  cones found in fovea.

5
Vision Physiology

• Retina LGN Primary Visual Cortex

From (2)
6
Landisman and Tso
From (8)
A Color selectivity map B Orientation map C
Color map with orientation pinwheels D
Orientation map with pinwheels and color map
overlaid
7
Landisman and Tso
From (8)

• How are these maps created in the cortex?
• How do the color, orientation, and O.D. maps
  interact?
• Landisman suggests ocular dominance is more
  related to color than orientation maps. True?

8
Center Surround Receptive Fields


9
Cone type distribution

• Random distribution of cone types. No S-cones in
  fovea.
• Ratio between L-M differs between patients 0.251
  to 91 without loss of color selectivity.
  (Brainard et al. 2000)

(7)
10
RGB Channels

• A single-wavelength source activates all cones.
• Multiple wavelengths can combine to duplicate
  single-wave stimulation.
• Light wavelength. (Cone mixture ratio)
• Intensity. (Spike frequency)
• Basic linear system to convert between inputs
  which allows us to do RGB monitors using three
  guns.
• Monitors can not do 100 of the viewable color
  spectrum.

11
RGB Channels
R
G
B
(6)
12
RGB Channels
R
Question Do the RGB channels match the
properties found in the biological cone
channels? Will it make a difference?
(6)
13
Optic nerve bandwidth constraints

• Optic nerve requires data compression to preserve
  visual information.
• Reduce redundant information
• Spiking frequency between cone-types are highly
  correlated because of cone sensitivity overlap.
• Spatially close locations have high color
  correlation.

14
Color Opponent Ganglion Cells
R- / G
R / G-

• Lateral Geniculate Neurons
• Concentric single-opponent (R/G)
• Concentric broad-band (GR)
• Not shown Co-extensive single-opponent B /
  (GR)-
• Cortical Neurons (not shown)
• Concentric double-opponent (Yellow/Blue,
  Red/Green contrasts)
• Complex double-opponent is similar to
  double-opponent but not as spatially selective.

G / R-
G- / R
(GR) (GR)-
(GR)- (GR)
15
RF-LISSOM Model
LISSOM ? Laterally Interconnected Synergetically
Self-Organizing Map
From (3)
16
LISSOM with LGN Layers

• Each Cell Type in the LGN represented with a
  separate Layer between Photoreceptors and V1.
• Each Layer has a receptive field with a possibly
  different Gaussian function.
• LISSOM normalizes and averages for different
  numbers of total Layers.

17
LISSOM Opponent-cell architecture
18
Problem with LISSOM?
Orientation Selectivity Map Baseline
Should be identical
Opponent Cell Control Stimulus (R/R)
19
Finally A Result. And it means?
Original Orientation Map Architecture
Opponent Cell Architecture with R/G
20
Initial observations

• Opponent networks are not as biased for certain
  directions as the built-in architecture.
• Biases are expected for natural images.
• Reason Combination of RFs?
• Maps not as selective in opponent-cell network.
• Reason Plotting? Combination of RFs?
• Fix?

21
Immediate steps

• Testing network for color selectivity neurons
  through LISSOM
• Find locations of neurons relative to orientation
  selectivity.

22
Landisman and Tso
From (8)

• Stay with uncontroversial R/G color opponent
  cells?
• No guarantee of results even at this level, too
  simple?
• How do double-opponent cells play a role?

23
Natural / Artificial Stimuli

• Natural
• Pro The Real Thing.
• Con What colors, objects in sample set?
• Artificial
• Pro Know what is learned.
• Con What is lost?

(6)
vs.
24
Future Directions? Feedback wanted.

• Stay with uncontroversial Red/Green color
  opponent cells?
• No guarantee of results even at this level, too
  simple?
• Orientation maps, motion/object detection?

25
References

• http//www.phys.ufl.edu/avery/course/3400/gallery
  
• http//webvision.med.utah.edu/Color.html
• Miikkulainen, Risto and Bednar, James A. and
  Choe, Yoonsuck and Sirosh, Joseph (1997)
  Self-Organization, Plasticity, and Low-level
  Visual Phenomena in a Laterally Connected Map
  Model of the Primary Visual Cortex, Psychology of
  Learning and Motivation, volume 36 Perceptual
  Learning, pp. 257-308.
• Kandel, Schwartz, and Jessell. Principles of
  Neural Science Third Edition 1991.
• James A. Bednar and Risto Miikkulainen (2003).
  Self-Organization of Spatiotemporal Receptive
  Fields and Laterally Connected Direction Maps,
  Neurocomputing 52-54473-480.
• http//www.brainfiber.com/flowers.jpg
• Lennie P. 2000. Color vision putting it
  together. Curr. Biol. 10(16)R589-91
• Landisman, Carole and Tso Daniel. Color
  Processing in the Macaque Striate Cortex
  Relationships to Ocular Dominance, Cytochrome
  Oxidase, and Orientation.